Introduction

This project aims to provide a better Chinese CLIP model. The training data used in this project consists of publicly accessible image URLs and related Chinese text descriptions, totaling 400 million. After screening, we ultimately used 100 million data for training. This project is produced by QQ-ARC Joint Lab, Tencent PCG. For more detailed information, please refer to the main page of the QA-CLIP project. We have also open-sourced our code on GitHub, QA-CLIP, and welcome to star!

Results

We conducted zero-shot tests on MUGE Retrieval, Flickr30K-CN, and COCO-CN datasets for image-text retrieval tasks. For the image zero-shot classification task, we tested on the ImageNet dataset. The test results are shown in the table below:

Flickr30K-CN Zero-shot Retrieval (Official Test Set):

Task	Text-to-Image			Image-to-Text
Metric	R@1	R@5	R@10	R@1	R@5	R@10
CN-CLIP_RN50	48.8	76.0	84.6	60.0	85.9	92.0
QA-CLIP_RN50	50.5	77.4	86.1	67.1	87.9	93.2
CN-CLIP_ViT-B/16	62.7	86.9	92.8	74.6	93.5	97.1
QA-CLIP_ViT-B/16	63.8	88.0	93.2	78.4	96.1	98.5
CN-CLIP_ViT-L/14	68.0	89.7	94.4	80.2	96.6	98.2
AltClip_ViT-L/14	69.7	90.1	94.8	84.8	97.7	99.1
QA-CLIP_ViT-L/14	69.3	90.3	94.7	85.3	97.9	99.2

MUGE Zero-shot Retrieval (Official Validation Set):

Task	Text-to-Image			Image-to-Text
Metric	R@1	R@5	R@10	R@1	R@5	R@10
CN-CLIP_RN50	42.6	68.5	78.0	30.0	56.2	66.9
QA-CLIP_RN50	44.0	69.9	79.5	32.4	59.5	70.3
CN-CLIP_ViT-B/16	52.1	76.7	84.4	38.7	65.6	75.1
QA-CLIP_ViT-B/16	53.2	77.7	85.1	40.7	68.2	77.2
CN-CLIP_ViT-L/14	56.4	79.8	86.2	42.6	69.8	78.6
AltClip_ViT-L/14	29.6	49.9	58.8	21.4	42.0	51.9
QA-CLIP_ViT-L/14	57.4	81.0	87.7	45.5	73.0	81.4

COCO-CN Zero-shot Retrieval (Official Test Set):

Task	Text-to-Image			Image-to-Text
Metric	R@1	R@5	R@10	R@1	R@5	R@10
CN-CLIP_RN50	48.1	81.3	90.5	50.9	81.1	90.5
QA-CLIP_RN50	50.1	82.5	91.7	56.7	85.2	92.9
CN-CLIP_ViT-B/16	62.2	87.1	94.9	56.3	84.0	93.3
QA-CLIP_ViT-B/16	62.9	87.7	94.7	61.5	87.6	94.8
CN-CLIP_ViT-L/14	64.9	88.8	94.2	60.6	84.4	93.1
AltClip_ViT-L/14	63.5	87.6	93.5	62.6	88.5	95.9
QA-CLIP_ViT-L/14	65.7	90.2	95.0	64.5	88.3	95.1

Zero-shot Image Classification on ImageNet:

Task	ImageNet
CN-CLIP_RN50	33.5
QA-CLIP_RN50	35.5
CN-CLIP_ViT-B/16	48.4
QA-CLIP_ViT-B/16	49.7
CN-CLIP_ViT-L/14	54.7
QA-CLIP_ViT-L/14	55.8

Getting Started

Inference Code

Inference code example：

from PIL import Image
import requests
from transformers import ChineseCLIPProcessor, ChineseCLIPModel

model = ChineseCLIPModel.from_pretrained("TencentARC/QA-CLIP-ViT-B-16")
processor = ChineseCLIPProcessor.from_pretrained("TencentARC/QA-CLIP-ViT-B-16")

url = "https://clip-cn-beijing.oss-cn-beijing.aliyuncs.com/pokemon.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
# Squirtle, Bulbasaur, Charmander, Pikachu in English
texts = ["杰尼龟", "妙蛙种子", "小火龙", "皮卡丘"]

# compute image feature
inputs = processor(images=image, return_tensors="pt")
image_features = model.get_image_features(**inputs)
image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True)  # normalize

# compute text features
inputs = processor(text=texts, padding=True, return_tensors="pt")
text_features = model.get_text_features(**inputs)
text_features = text_features / text_features.norm(p=2, dim=-1, keepdim=True)  # normalize

# compute image-text similarity scores
inputs = processor(text=texts, images=image, return_tensors="pt", padding=True)
outputs = model(**inputs)
logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
probs = logits_per_image.softmax(dim=1)

Acknowledgments

The project code is based on implementation of Chinese-CLIP, and we are very grateful for their outstanding open-source contributions.